AI can write a reassuring medical message. That does not mean it feels concern. A 2023 study found that blinded clinicians preferred chatbot answers to...
AI-Written Doctor Messages: Can a Machine Sound Empathetic Without Caring?
AI can write a reassuring medical message. That does not mean it feels concern. A 2023 study found that blinded clinicians preferred chatbot answers to physician replies for nearly 80% of patient questions, and rated them empathetic almost ten times as often. The result is striking, but it does not settle what patients deserve from their doctors.
Key Takeaways- Fluent, considerate wording is not the same as felt empathy or personal responsibility.
- A 2023 comparison favored chatbot responses, but its setup did not recreate an ongoing doctor-patient relationship.
- AI can help clinicians communicate more clearly if a clinician checks, edits, and owns each message.
- Patients should be told when AI materially contributes to their care, especially when it shapes clinical advice.
The number is hard to ignore. In the study, researchers compared physician and chatbot answers to 195 questions posted on a public online forum; blinded clinicians favored chatbot answers nearly four times out of five, and rated them empathetic almost ten times as often. Yet the comparison had limits: volunteer physicians typed brief replies on Reddit, not considered messages to their own patients. That distinction matters. A public answer judged on its wording is not the same thing as care delivered over time, with a known history, consequences, and a clinician who can be held accountable.
I read the result as a challenge to medical practice, not proof that software cares better than people. AI can help a tired clinician find plain, kind words after a long clinic day. But a patient’s need is not only for a pleasant sentence. It is also to know that someone has listened, understood the risk, and will take responsibility for what happens next.
What Is AI Empathy in Healthcare?
AI empathy in healthcare is the use of software to produce language that recognizes a person’s feelings, responds with warmth, or communicates concern. That definition describes the words on the screen. It does not establish that the software feels anything. Today’s language models generate text by identifying patterns and predicting likely sequences; they can imitate the outward shape of empathy without having human emotion, personal relationships, or moral responsibility.
That distinction can sound philosophical. It becomes practical when an AI drafts a portal message about a worrying test result, a new diagnosis, or a patient who says pain is getting worse. A sentence such as “I’m sorry this has been frightening” may feel meaningful to the reader. But if the clinician has not reviewed the case, the next steps, and the wording, the warmth can cover a serious gap: nobody has actually taken ownership of the message.
The 2023 study in JAMA Internal Medicine offers a useful example of what AI can do—and what its result cannot prove. Researchers assessed answers to 195 patient questions from a public forum. Blinded clinicians preferred chatbot-generated answers in nearly 80% of comparisons and judged them empathetic almost ten times as often as physician responses. The researchers measured the perceived quality of written replies. They did not measure whether patients felt cared for over months, whether the replies improved outcomes, or whether the chatbot could stand behind its advice when circumstances changed. The published study should be read with those boundaries in view.
A further limit is the setting. The physician answers were short volunteer replies on Reddit, not necessarily the sort of communication a doctor would write to a patient whose story, preferences, and medical record they knew. This is not a reason to dismiss the finding. It is a reason not to inflate it. A chatbot can produce polished prose; a clinician’s duty includes judgment, follow-up, confidentiality, and accountability.
The philosopher Emmanuel Levinas wrote about the ethical force of encountering another person face to face. In medicine, that encounter need not happen in the same room: a portal message can still carry attention and responsibility. But a face, a voice, and a caring phrase are not interchangeable. AI may provide words; the clinician must supply the human commitment behind them.
Why It Matters: Benefits and Limits
Used carefully, AI-assisted writing can help. Used carelessly, it can make a thin interaction sound more caring than it is. Frankly, both possibilities deserve more attention than claims that a chatbot is simply “better at empathy.”
Potential benefits include:
- Clearer explanations. A clinician can ask software to turn technical language into a more readable draft. That may help patients understand what a result means, what remains uncertain, and what action to take. The clinician still needs to verify every clinical claim.
- A more considerate tone. When a busy doctor is exhausted or pressed for time, a draft can suggest a calmer opening or a less abrupt explanation. A useful prompt might ask for plain language and a respectful tone, not for the machine to pretend it has personal feelings.
- Faster routine communication. Templates and carefully supervised drafting may reduce repetitive typing. Time saved is valuable only if it is used well—perhaps to answer a complex concern, review a chart, or speak with a patient who needs more than a standard reply.
- Support for accessibility. Drafts can be adjusted to shorter sentences or a lower reading level. Clinicians must still consider language access, disability, cultural context, and whether the message fits the individual.
The strongest case for AI is practical assistance, not replacement. A writing tool can offer an initial draft, but it cannot know whether the patient’s apparent question is masking fear, whether a new symptom needs urgent assessment, or whether a previous conversation changed the plan unless it has the relevant information—and even then, the clinician must interpret it. A warm sentence that misses a red flag is not good care.
There are also risks. A model can state something inaccurate with confidence, omit an important warning, or produce a tone that seems patronizing. It may include personal details in a third-party system if the clinic has not set appropriate privacy protections. And when software generates a message that sounds personal, the patient may reasonably assume the doctor chose each word. Concealing meaningful AI use can weaken trust if the patient later learns otherwise.
The ethical standard is not that every message must be handwritten. Medicine has always used tools, staff, templates, and dictation. The standard is that a patient’s dignity and welfare remain central—and that a responsible human reviews and owns communication that affects care. As the World Health Organization’s guidance on AI for health emphasizes, human oversight, transparency, and accountability matter when AI is used in health settings. The tool should serve the relationship, not impersonate it.
How to Use AI-Drafted Patient Messages Responsibly
A safe workflow is not complicated, though it does require discipline. I would treat AI as a junior drafting assistant: useful for a first pass, never the final authority. Here is a practical approach for a clinic or individual clinician.
- Decide what the tool is allowed to do. Set written rules before use. Is it approved to improve grammar, summarize a clinician’s instructions, or draft routine appointment information? Is it prohibited from independently interpreting a test, recommending a treatment change, or replying to urgent symptoms? Boundaries reduce the temptation to use a general writing tool as an untested clinical decision-maker.
- Protect patient information. Use only systems approved by the health organization, with appropriate security and privacy review. Do not paste identifiable patient details into a consumer tool unless the organization has confirmed that use is permitted and safe. Remove unnecessary identifiers when possible. Privacy is not a footnote; patient trust depends on it.
- Give the model a narrow, factual task. Provide the relevant clinician-approved information and ask for a draft that is clear, concise, and respectful. Do not ask it to invent a diagnosis or fill gaps with guesses. If the model lacks information, the right answer is to ask the clinician or patient—not to improvise.
- Check the clinical substance first. Confirm names, dates, medications, doses, test results, follow-up timing, and instructions. Check that the response addresses the actual question and does not leave out a warning sign or an action the patient needs to take. A grammatical sentence can still be medically wrong.
- Read for the human meaning. Ask whether the reply recognizes what the patient is worried about, avoids false reassurance, and explains uncertainty honestly. A message can be warm without claiming “I know exactly how you feel.” If it promises a call, a referral, or a result review, confirm that someone will do it.
- Edit the draft into a message you can stand behind. Replace generic expressions with relevant details where appropriate. Remove phrases that sound intimate but are not true of the clinician’s relationship with the patient. The final wording should reflect the doctor’s judgment and the actual plan, not merely the model’s idea of a comforting reply.
- Make responsibility visible. The clinician who sends or approves a message remains accountable for its content. If an AI tool contributed substantially, follow organizational policy on disclosure. For sensitive or high-stakes communication, explain the process plainly and invite questions. The American Medical Association’s guidance on augmented intelligence frames AI as assistance to human intelligence, not a substitute for professional judgment.
- Create a path for escalation. A portal message should tell patients what to do if symptoms worsen, when to seek urgent help, and how to reach a person. Do not let an automated reply become a locked door. Review missed messages and response times, and make sure a human can step in when the conversation becomes complex.
- Audit results with patients and staff. Track errors, complaints, delayed responses, and differences in performance across patient groups. Ask patients whether the messages are understandable and whether they know who is responsible for their care. If the tool creates extra correction work or erodes confidence, change the process or stop using it. Convenience is not evidence of benefit.
The point is not to ban assistance. It is to preserve a clear line between drafting and deciding, between pleasant phrasing and genuine clinical accountability. The best workflow keeps the machine in the background and the patient’s actual needs in view.
AI-Assisted Messages vs. Human-Written Replies
These approaches are not true opposites. A clinician can write a message without AI, or use AI for a draft and then review it. The meaningful comparison is between a machine-generated response presented as if it were complete and a response for which a clinician accepts responsibility. The table summarizes the evidence and the trade-offs.
| Measure | AI-generated or AI-assisted reply | Human-written physician reply |
|---|
| Wording in the 2023 study | Blinded clinicians preferred chatbot answers in nearly 80% of comparisons and rated them empathetic almost ten times as often. | Volunteer physicians wrote brief replies to questions posted on a public forum. The study did not compare long-term care relationships. |
| What it can do well | Produce fluent drafts, explain concepts in plain language, and suggest a more considerate tone. | Apply clinical context, understand a patient’s history, recognize responsibility, and respond to nuance. |
| Main limitation | It can sound caring without feeling concern, and may produce inaccurate or incomplete text. | People can be rushed, unclear, or emotionally depleted; human authorship alone does not guarantee empathy. |
| Accountability | The software cannot personally answer for harm or follow through on a promise. | A clinician can be identified, questioned, and held professionally responsible for an approved message. |
| Best role | A supervised writing aid with privacy safeguards and human review. | The accountable source of clinical judgment and the relationship with the patient. |
The study’s result is about how reviewers rated particular answers, not a verdict on every doctor or every AI system. It should prompt clinicians to ask why some generated replies seemed more empathetic. Perhaps the models used more words, acknowledged distress directly, or organized information more clearly. Those are teachable communication skills. They can improve human replies without pretending that software has feelings.
The comparison also exposes an uncomfortable truth: patients may receive rushed, terse answers from real clinicians. That deserves attention in its own right. If an AI draft improves readability, the clinic should ask whether it has also made time for a human to understand the concern. Better wording matters, but it cannot compensate for an understaffed inbox, poor follow-up, or a clinician who never checks the message.
Common Mistakes to Avoid
Mistake 1: Treating empathetic wording as proof of empathy. A message may contain apologies, validation, and reassurance while missing the patient’s actual fear. These phrases are not automatically harmful, but they become hollow when they are generic or disconnected from action. “I’m sorry this is difficult” should lead into a relevant answer, not serve as decorative cover.
Mistake 2: Believing a high preference score proves better care. The 2023 study compared answers to forum questions and asked clinicians to rate them. It did not test clinical outcomes, patient trust over time, or the quality of a doctor’s continuing relationship with a patient. A promising comparison is a reason to study a tool further, not permission to declare the question settled.
Mistake 3: Sending drafts without checking the medicine. Models can confuse details, leave out context, or sound certain where the evidence is uncertain. A clinician should not skim for tone alone. Check the facts, the plan, the safety instructions, and whether the patient’s question was answered. If something is unclear, pause and investigate.
Mistake 4: Using warmth to avoid a difficult conversation. Some news needs a call or a scheduled discussion, not a polished portal paragraph. A patient facing a serious diagnosis may need time to ask questions and hear uncertainty explained. The tool should not be used to make difficult news easier for the clinic to send while leaving the patient alone with it.
Mistake 5: Hiding how the message was produced. Not every spellcheck or formatting aid needs an announcement. But if AI substantially shapes advice or communication, clinics should have a transparent policy and follow it. Disclosure is especially important when a patient asks, when the tool influences a decision, or when the system’s role could reasonably affect trust.
Mistake 6: Assuming human-only means humane. People can be brusque, exhausted, or unclear. Doctors need time, training, and working systems that make careful communication possible. AI may help with some burdens, but it should not become an excuse to tolerate poor staffing or to remove the human contact patients need.
Mistake 7: Forgetting who bears the consequences. The software does not worry about a missed diagnosis, an incorrect dose, or a patient who follows a bad instruction. The person who sends the message and the organization that deploys the tool must have a way to correct errors and respond to harm. The AMA’s physician-patient relationship ethics guidance is a useful reminder that professional care involves obligations, not just well-phrased exchanges.
Let’s be real: a portal message is part of care, not mere clerical output. The body and life described in that inbox belong to a person whose dignity deserves attention. Good technology should help clinicians honor that responsibility, not make a simulation of concern look like the whole job.
Frequently Asked Questions
Can AI feel empathy toward a patient?
Current AI language tools can produce words that people perceive as empathetic, but they do not have human feelings, personal concern, or a relationship with the patient. It is more accurate to say they can imitate empathetic communication than to say they feel empathy. Whether a message helps a patient also depends on its accuracy, context, and the care that follows.
What did the 2023 JAMA Internal Medicine study find?
Researchers compared physician and chatbot answers to 195 patient questions posted on a public online forum. Blinded clinicians preferred chatbot responses nearly 80% of the time and rated them empathetic almost ten times as often. The study assessed written responses in a limited setting; it did not show that chatbots provide better ongoing care or have more genuine concern than physicians.
Is it safe for doctors to use AI to draft portal messages?
It can be appropriate when the tool has been reviewed and approved, patient information is protected, and a clinician checks the final message for accuracy, completeness, tone, and safe next steps. AI should not independently diagnose, make high-stakes decisions, or send unreviewed clinical advice. Each health organization needs clear rules and a way to escalate complex or urgent concerns.
Should a doctor tell a patient when AI helped write a message?
The answer depends on the organization’s policy and how much the system contributed. A clinic should be transparent about meaningful AI use, especially if it shaped clinical advice or if a patient asks. Whatever the disclosure rule, the clinician must not imply that the software itself understands the patient or can take responsibility for the care plan.
Final Thought
The most important question is not whether a machine can produce a sentence that sounds kind. The study suggests it can. The harder question is whether the person receiving that sentence is being heard, protected, and cared for by someone who will answer for what the message says. Smooth language is easy to mistake for attention, especially when patients are worried and the health system is busy.
Cyrano de Bergerac gives us a fitting image: one character supplies the words while another receives the admiration. In clinical communication, the danger is not simply that AI writes a draft. Ghostwriting has existed for centuries, and tools have always shaped how professionals communicate. The danger is that the human being who carries the duty may fade from view—that a patient is left with eloquent words but no clear person to turn to when the advice fails, the symptoms change, or the fear remains.
A clinician can use AI and still communicate honestly. That takes more than clicking “accept.” It takes careful review, an accurate plan, privacy safeguards, transparent practice, and a willingness to respond when the patient needs something beyond a draft. It also takes the humility to admit that tired humans sometimes communicate poorly and the discipline to use technology to improve—not disguise—that weakness.
Our bodies and our lives are not entries in an inbox. They deserve care rooted in respect, judgment, and responsibility. Let AI help find the words when it is useful; let a real person mean them, verify them, and stand behind them.